Re: [PATCH 2.6.12-rc4] cpuset exit NULL dereference fix

[Paul Jackson <pj@sgi.com>]

Simon wrote:
> > Would it make sense, Simon, to recommend to Andrew that
> > he take the simple patch I submitted yesterday ...
> > 
> > Then, when we understand ... offer up a second patch?
> 
> Of course !

Ok - I'll resubmit that patch.  Hopefully you can reply to that
resubmitted patch with an "Acked-by: ..."

Andrew withdrew the original patch, when it became a matter needing
further discussion.

> My point is only that if you think there is a scaling problem in 
> taking cpuset_sem for each call to cpuset_exit(), that scaling problem 
> won't disappear by taking cpuset_sem only for 'notify_on_remove' cpusets, 

Yes - that is a good and valid point.

I also lack any real evidence of a scaling problem.  It's just a
theoretical concern.  My unreliable weather forecast is that it will be
a while before it's a serious concern.

This means I am willing to take simple measures to minimize the concern,
but I'd prefer to await hard evidence of the problem before trying more
ambitious measures.

==

My impression is that cpusets has two classes of users:
 1) Extreme HPC apps, scaling to hundreds or thousands of CPUs, and
 2) More mixed or service oriented apps, with less extreme scaling.

I also suspect that it is the second class that most requires
notify_on_release.  The extreme HPC guys have less need for
notify_on_release.

So I find a solution that lets one trade off extreme scaling versus
heavier use of notify_on_release to be appealing, if it can be done
trivially, as this patch does.

-- 
                  I won't rest till it's the best ...
                  Programmer, Linux Scalability
                  Paul Jackson <pj@engr.sgi.com> 1.650.933.1373, 1.925.600.0401